Generative Vision Atlas

Research line · ascendant

Training-free attention manipulation

Edit by intervening in attention maps and features at sampling time, with no training and often no change to the latent at all.

What defines membership

A pretrained generator already contains the structure of the source image in its activations; editing is a matter of reaching in and holding the right parts fixed.

How the line developed

Read top to bottom: what came before, the idea itself, the evidence for it, what improved, and where it breaks.

The idea

Prompt-to-Prompt · 2022-08landmark

Cross-attention maps bind words to regions, so recording and reinjecting them edits the prompt without touching the model.

Improvement

Plug-and-Play · 2022-11core

Injects spatial features and self-attention maps from an inverted guidance image into specific decoder layers.

MasaCtrl · 2023-04core

Mutual self-attention retrieves content from the source branch rather than resampling it, handling non-rigid edits.

Stable Flow · 2024-11strong-followup

Finds the few vital layers that drive image formation in a diffusion transformer, since transformers lack the UNet hierarchy.

KV-Edit · 2025-02core

Freezes background key-value pairs so unedited content never re-enters generation — the only architecturally earned preservation guarantee in this section.

What it gets right

  • No training cost, so it applies immediately to any released model
  • Provides the strongest preservation guarantees available, by not regenerating the unedited content at all
  • Interpretable: the intervention point is an identifiable layer or attention map

Where it is weak

  • Intervention points must be rediscovered for each new architecture
  • Bounded by what the base model can already produce
  • Struggles when the edit requires the whole scene to change

Open problems it has not solved